Data processing method and device, electronic equipment and storage medium

Abstract

The application relates to a data processing method and device, electronic equipment and a storage medium. The method comprises the following steps: obtaining a first target object and a second candidate object, the second candidate object being a recommended object of the first target object; taking the first target object and the second candidate object as inputs, and obtaining a first correlation relationship prediction result by using a target graph neural network, the first correlation relationship prediction result indicating a correlation relationship between the first target object and the second candidate object; and in the case that the first correlation relationship prediction result meets preset requirements, recommending the second candidate object to the first target object. The application improves the accuracy and efficiency of the correlation relationship prediction between the first target object and the second candidate object, and further improves the accuracy of object recommendation. The embodiments of the application can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation and intelligent entertainment.

Description

Technical Field

[0001] This application relates to the field of Internet communication technology, and in particular to a data processing method, apparatus, electronic device and storage medium. Background Technology

[0002] With the development of internet communication technology, various internet products have emerged, and the user experience they provide is constantly improving. For example, the recommended objects (such as multimedia objects and product objects) are becoming increasingly accurate. In related technologies, popular objects associated with the user's input or selection are often recommended. However, this method of determining the objects to be recommended has limitations; it doesn't consider the relationship between the recommended objects and the user, which can easily affect the accuracy of the recommendations. Therefore, a more accurate and effective object recommendation solution is needed. Summary of the Invention

[0003] To address at least one of the aforementioned technical problems, this application provides a data processing method, apparatus, electronic device, and storage medium:

[0004] According to a first aspect of this application, a data processing method is provided, the method comprising:

[0005] Obtain a first target object and a second candidate object, wherein the second candidate object is the object to be recommended from the first target object;

[0006] Using the first target object and the second candidate object as input, a first association prediction result is obtained using a target graph neural network. The first association prediction result indicates the association relationship between the first target object and the second candidate object.

[0007] If the first correlation prediction result meets the preset requirements, the second candidate object is recommended to the first target object;

[0008] The target graph neural network is obtained by adversarial training of a preset graph neural network and a generator network based on a first association graph and a second association graph. The input of the preset graph neural network is the first association graph and the second association graph. The second association graph is output by the generator network based on the first association graph. The first association graph includes multiple first sample objects and multiple second sample objects as nodes and edges between connected nodes. The edges represent the association relationship between the connected nodes.

[0009] According to a second aspect of this application, a data processing apparatus is provided, the apparatus comprising:

[0010] Acquisition module: used to acquire a first target object and a second candidate object, wherein the second candidate object is the object to be recommended from the first target object;

[0011] Prediction module: used to obtain a first association prediction result by using the first target object and the second candidate object as input and a target graph neural network, wherein the first association prediction result indicates the association relationship between the first target object and the second candidate object;

[0012] Recommendation module: used to recommend the second candidate object to the first target object when the prediction result of the first association relationship meets the preset requirements;

[0013] The target graph neural network is obtained by adversarial training of a preset graph neural network and a generator network based on a first association graph and a second association graph. The input of the preset graph neural network is the first association graph and the second association graph. The second association graph is output by the generator network based on the first association graph. The first association graph includes multiple first sample objects and multiple second sample objects as nodes and edges between connected nodes. The edges represent the association relationship between the connected nodes.

[0014] According to a third aspect of this application, an electronic device is provided, the electronic device including at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores at least one instruction or at least one program, the at least one instruction or at least one program being loaded and executed by the at least one processor to implement the data processing method as described in the first aspect.

[0015] According to a fourth aspect of this application, a computer-readable storage medium is provided, wherein at least one instruction or at least one program is stored therein, the at least one instruction or at least one program being loaded and executed by a processor to implement the data processing method as described in the first aspect.

[0016] According to a fifth aspect of this application, a computer program product is provided, the computer program product comprising at least one instruction or at least one program segment, the at least one instruction or at least one program segment being loaded and executed by a processor to implement the data processing method as described in the first aspect.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application.

[0018] Implementing this application will have the following beneficial effects:

[0019] This application utilizes a target graph neural network to obtain the first association prediction result, improving the accuracy and efficiency of predicting the association relationship between the first target object and the second candidate object, thereby improving the accuracy of object recommendation for the first target object based on the first association prediction result. The training samples are association graphs. Learning from these graphs helps the first association prediction result determine whether there is an association between the first target object and the second candidate object, thus improving the object recommendation effect. Simultaneously, the acquisition of the target graph neural network involves adversarial training with a generator network. The generator network is responsible for providing adversarial examples. This training process improves the robustness and generalization ability of the target graph neural network model.

[0020] Other features and aspects of this application will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0021] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This diagram illustrates an application environment according to an embodiment of the present application.

[0023] Figure 2 A flowchart illustrating a data processing method according to an embodiment of this application is shown;

[0024] Figure 3 This diagram illustrates the process of training a target graph neural network according to an embodiment of this application.

[0025] Figure 4 A flowchart illustrating the process of obtaining a second association graph from an augmented graph according to an embodiment of this application is shown.

[0026] Figure 5 This diagram illustrates a flowchart of parameter tuning for the discrimination network and the generation network according to an embodiment of this application.

[0027] Figure 6 It also shows a schematic diagram of the process of training a target graph neural network according to an embodiment of this application;

[0028] Figure 7 This diagram illustrates a device block diagram according to an embodiment of the present application;

[0029] Figure 8 A schematic diagram of an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0032] Various exemplary embodiments, features, and aspects of this application will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0033] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0034] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0035] Furthermore, to better illustrate this application, numerous specific details are provided in the following detailed description. Those skilled in the art should understand that this application can be implemented without certain specific details. In some instances, methods, means, components, and circuits well-known to those skilled in the art have not been described in detail in order to highlight the main points of this application.

[0036] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0037] Graph Neural Networks (GNNs) are a general term for algorithms that use neural networks to learn graph-structured data, extract and discover features and patterns in graph-structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. Currently, representative GNNs include Graph Attention Networks (GATs) and Lightweight Graph Convolutional Networks (LightGCNs).

[0038] DA (Data Augmentation): Data augmentation.

[0039] A bipartite graph, also known as a bipartite graph, is a special model in graph theory. Let G = (V, E) be an undirected graph. If a vertex V can be partitioned into two disjoint subsets (A, B), and each edge (i, j) in the graph is associated with two vertices i and j belonging to these two distinct vertex sets (i in A, j in B), then graph G is called a bipartite graph.

[0040] Multilayer Perceptron (MLP): A feedforward artificial neural network model that maps multiple input datasets to a single output dataset.

[0041] The Hadamard product is a type of matrix operation.

[0042] The Gumbel-Max trick is essentially a technique based on reparameters, which transforms sampling from a categorical distribution into finding the maximum value among k values.

[0043] BPR (Bayesian Personalized Ranking): Bayesian personalized ranking.

[0044] NCE (Noise Contrastive Estimation): Noise Contrastive Estimation.

[0045] Please see Figure 1 , Figure 1This diagram illustrates an application environment according to an embodiment of the present application. The application environment may include a client 10 and a server 20. The client 10 and server 20 can be directly or indirectly connected via wired or wireless communication. Users experience relevant business services provided by the target application through the client 10. The server 20 provides background services related to the target application to the client 10. Taking a user's account as an example, the server 20 obtains the first target object and a second candidate object, where the second candidate object is the object to be recommended by the first target object. Then, using the first target object and the second candidate object as input, a target graph neural network is used to obtain a first association prediction result, which indicates the association between the first target object and the second candidate object. Finally, if the first association prediction result meets preset requirements, the second candidate object is recommended to the first target object. It should be noted that... Figure 1 This is just one example.

[0046] The client can be a physical device such as a smartphone, computer (e.g., desktop computer, tablet, laptop), augmented reality (AR) / virtual reality (VR) device, digital assistant, smart voice interaction device (e.g., smart speaker), smart wearable device, smart home appliance, in-vehicle terminal, etc., or software running on the physical device, such as a computer program. The operating system corresponding to the client can be Android, iOS (a mobile operating system developed by Apple), Linux, Microsoft Windows, etc.

[0047] The server side can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The server may include network communication units, processors, and memory, etc. The server side can provide backend services to the corresponding clients.

[0048] In practical applications, the client 10 can independently perform the above data processing steps on the first target object ("obtain the first target object and the second candidate object" to "recommend the second candidate object to the first target object if the prediction result of the first association meets the preset requirements"), or the client 10 and the server 20 can interact to perform the above data processing steps on the first target object.

[0049] The data processing scheme provided in this application can utilize technologies related to Computer Vision (CV) and Natural Language Processing (NLP). Computer vision is a science that studies how to enable machines to "see." More specifically, it refers to using cameras and computers to replace human eyes for target recognition, measurement, and other machine vision tasks, and further performing image processing to create images more suitable for human observation or transmission to instruments for detection. As a scientific discipline, computer vision studies related theories and technologies, attempting to establish artificial intelligence systems capable of extracting information from images or multidimensional data. Computer vision technologies typically include image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, 3D object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping (SLAM), autonomous driving, intelligent transportation, and other technologies, as well as common biometric recognition technologies such as facial recognition and fingerprint recognition.

[0050] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language people use in daily life—and thus it has a close relationship with linguistic research. NLP technologies typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0051] It should be noted that when the first target object, second candidate object, first association diagram and second association diagram, etc., which are related to user information, are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0052] Figure 2 This diagram illustrates a flow chart of a data processing method according to an embodiment of this application, such as... Figure 2 As shown, the method includes:

[0053] S201: Obtain a first target object and a second candidate object, wherein the second candidate object is the object to be recommended by the first target object;

[0054] In this embodiment, the server or client obtains a first target object and a second candidate object. The first target object and the second candidate object are two types of objects. For example, the first target object and the second candidate object are multimedia objects (such as text, audio, video, etc.); the first target object is an account object, and the second candidate object is a product object. For account objects, it can also include subcategories. Accordingly, the first target object can indicate a job-seeking account (or a recruitment account), and the second candidate object can indicate a recruitment account (or a job-seeking account).

[0055] The second candidate object is the object to be recommended from the first target object. This can be understood as follows: 1) The second candidate object can be an object that already exists within a preset time period. The preset time period can indicate all or part of the historical time related to the first target object (such as the most recent month, the most recent six months, etc.). Within the preset time period, there is no association between the first target object and the second candidate object. Whether there is an association between them can be defined by whether there is interaction between the first target object and the second candidate object. Here, interaction can refer to the interactive behavior performed by the first target object on the second candidate object. Interactive behavior can include, but is not limited to, clicking, browsing, listening, watching, collecting, following, adding friends, sending messages, downloading, forwarding, and sharing. Of course, the interactive behavior here can be more finely defined based on historical feedback and business scenarios. 2) The second candidate object can also be a newly added object.

[0056] There can be at least one second candidate object. At least one second candidate object can be determined from the candidate object pool. The candidate object pool includes multiple candidate objects that satisfy the limiting conditions for "second candidate object" in 1) and / or 2) above. At least one second candidate object can be a subset of candidate objects randomly determined from the candidate object pool, or it can be a subset of candidate objects with higher current popularity determined from the candidate object pool.

[0057] S202: Using the first target object and the second candidate object as input, a first association prediction result is obtained using a target graph neural network. The first association prediction result indicates the association relationship between the first target object and the second candidate object.

[0058] In this embodiment, the server or client takes a first target object and a second candidate object as input and uses a target graph neural network to obtain a first association prediction result. The target graph neural network can predict whether an association exists between the first target object and the second candidate object. That is, the first association prediction result can indicate whether an association exists between the first target object and the second candidate object, or whether no association exists. The target graph neural network can also predict the probability that an association exists between the first target object and the second candidate object. In other words, the first association prediction result can indicate the probability that an association exists between the first target object and the second candidate object.

[0059] The target graph neural network is obtained by adversarially training a pre-defined graph neural network and a generator network based on a first and a second association graph. The input to the pre-defined graph neural network is the first and second association graphs, and the second association graph is the output of the generator network based on the first association graph. In adversarial training, the pre-defined graph neural network learns from the first and second association graphs. The first association graph serves as the real sample, and the second association graph serves as the adversarial sample. The training objective of the pre-defined graph neural network is to distinguish between real and adversarial samples in order to better capture information from real samples. The generator network is responsible for providing adversarial samples, and its training objective is to provide adversarial samples that are difficult for the pre-defined graph neural network to distinguish. The information captured from real samples may include at least one of the following: information related to the formation of true edges (i.e., edges between connected nodes), such as the characteristics of each connected node and the feature relationships between connected nodes; and information related to the formation of pseudo edges (i.e., virtual edges between unconnected nodes), such as the characteristics of each unconnected node and the feature relationships between unconnected nodes.

[0060] The processing of the first target object and the second candidate object by the target graph neural network can be understood as follows: 1) If the input objects are a pair of first target objects and second candidate objects, the target graph neural network determines whether the first target object and the second candidate object can participate in the formation of "true edges", or what probability they have of participating in the formation of "true edges"; 2) If the input objects are a relationship to be processed consisting of the first target object and the second candidate object, the relationship to be processed can be regarded as an edge to be processed, and the target graph neural network determines whether the edge to be processed is a "true edge", or what probability it is a "true edge".

[0061] The first relationship diagram is a bipartite diagram, which can be used as a reference. Figure 6The first relationship graph includes multiple first sample objects and multiple second sample objects as nodes, and edges between connected nodes, where edges represent relationships between connected nodes. A connected node represents a connected first sample object and a connected second sample object. First sample objects and second sample objects are two types of objects. The aforementioned first target object and the first sample object are of the same type. The first target object can be any object among the multiple first sample objects, or it can be a similar object to any object among the multiple first sample objects. The aforementioned second candidate object and the second sample object are of the same type. The second candidate object can be any object among the multiple second sample objects, or it can be a similar object to any object among the multiple second sample objects. Multiple second sample objects can be determined through the following steps: for each first sample object, determine at least one heterogeneous object that has a relationship with the first sample object; then, based on the at least one heterogeneous object corresponding to each first sample object, obtain multiple second sample objects. The existence of a relationship between first sample objects and heterogeneous objects can be defined by whether there is interaction between them. The interaction here refers to the interactive behavior performed by the first sample object on a different type of object. This interaction behavior can include, but is not limited to, clicking, browsing, listening, watching, saving, following, downloading, forwarding, and sharing. Of course, the interactive behavior can be further refined based on historical feedback and business scenarios. It should be noted that the interactive behaviors (such as clicking) between the first and second sample objects in the first relationship graph are generally the same as the interactive behaviors (such as clicking) not involved between the first target object and the second candidate object.

[0062] The second relationship diagram is also a bipartite diagram, which can be used as a reference. Figure 6 The second association graph originates from the first association graph. It can be understood that the second association graph also includes the aforementioned multiple first sample objects and multiple second sample objects as nodes, but the connected nodes differ from those in the first association graph. Correspondingly, the edges between the connected nodes in the second association graph also differ. The differences between the connected nodes in the second association graph and those in the first association graph can be manifested as follows: 1) The connected nodes in the second association graph are part of the connected nodes in the first association graph; 2) The connected nodes in the second association graph involve new connected nodes that do not belong to the connected nodes in the first association graph; 3) 1) + 2); 4) The connected nodes in the second association graph are all the connected nodes in the first association graph + 2). For example, the first association graph includes first sample object 1-2 and second sample object ac, with connected nodes 1-a, 1-b, and 2-c. The second association graph includes first sample object 1-2 and second sample object ac, with connected nodes 1-b and 1-c.

[0063] S203: If the first association prediction result meets the preset requirements, recommend the second candidate object to the first target object;

[0064] In this embodiment of the application, if the prediction result of the first association relationship meets the preset requirements, the server or the client recommends a second candidate object to the first target object.

[0065] If a target graph neural network is used to predict whether a relationship exists between the first target object and the second candidate object, and the prediction result of the first relationship indicates that a relationship exists between the first target object and the second candidate object, it can be determined that the prediction result of the first relationship meets the preset requirements, and the second candidate object is recommended to the first target object. If a target graph neural network is used to predict whether a relationship exists between the first target object and the second candidate object, and the prediction result of the first relationship indicates that a relationship does not exist between the first target object and the second candidate object, it can be determined that the prediction result of the first relationship does not meet the preset requirements, and the second candidate object is not recommended to the first target object.

[0066] If a target graph neural network is used to predict the probability that there is a correlation between a first target object and a second candidate object, a lower probability limit should be pre-defined. If the probability of a correlation between the first target object and the second candidate object, as indicated by the first correlation prediction result, is greater than this lower probability limit, the second candidate object is recommended to the first target object; otherwise, the second candidate object is not recommended to the first target object.

[0067] Target graph neural networks (Graph Neural Networks) are trained models with high generalization ability. Using Graph Neural Networks for association prediction can improve the adaptability and reliability of association prediction. The model training is continuous, and the Graph Neural Network is constantly updated. It can be understood that the Graph Neural Network used for association prediction can be the result of the previous training or the basis for the next training. The training process of the model will be described below:

[0068] like Figure 3 , 6 As shown, the preset graph neural network is a discriminative network, and the target graph neural network is obtained through the following training method:

[0069] S301: Obtain the first association graph;

[0070] S302: Perturb the first association graph using the generator network to obtain the second association graph;

[0071] S303: The discriminant network is used to perform edge discrimination on the first association graph and the second association graph respectively to obtain a first discrimination result corresponding to the first association graph and a second discrimination result corresponding to the second association graph;

[0072] S304: Based on the first discrimination result and the second discrimination result, the parameters of the discrimination network and the generator network are updated using the target loss function;

[0073] S305: Repeat the above steps of obtaining the second association graph and updating the parameters of the discriminant network and the generator network until the value of the target loss function satisfies the convergence condition, and determine that the discriminant network corresponding to the current parameters is the target graph neural network.

[0074] In adversarial training, the generator network and the discriminator network learn in a competitive and adversarial manner, which can improve the robustness of the target graph neural network. The input of the generator network t is a first association graph G, and the output of the generator network is a second association graph t(G). The second association graph is the result of perturbing the first association graph. The discriminator network f performs edge discrimination on the first association graph G, and the first discrimination result f(G) is used to indicate the predicted true edges and predicted false edges in the first association graph. The discriminator network f performs edge discrimination on the second association graph t(G), and the second discrimination result f(t(G)) is used to indicate the predicted true edges and predicted false edges in the second association graph.

[0075] For the first discrimination result, the first association graph includes actual true edges and actual false edges. A predicted true edge is not necessarily an actual true edge, and a predicted true edge may be an actual false edge. A predicted false edge is not necessarily an actual false edge, and a predicted false edge may be an actual true edge.

[0076] Regarding the second discrimination result, the second association graph includes perturbed true edges and perturbed pseudo edges. A predicted true edge is not necessarily a perturbed true edge; a predicted true edge may be a perturbed pseudo edge. Conversely, a predicted pseudo edge is not necessarily a perturbed pseudo edge; a predicted pseudo edge may be a perturbed true edge. It should be noted that the perturbed true edge may differ from the actual true edge mentioned above, and the perturbed pseudo edge may differ from the actual pseudo edge mentioned above.

[0077] Since the discriminative network is trained to distinguish between real and adversarial samples, and the generative network is trained to provide adversarial samples that are difficult for the predefined graph neural network to distinguish, if the first discrimination result is better than the first predefined standard in predicting actual real edges as real edges and actual pseudo edges as pseudo edges, then the discriminative network has a good discrimination effect. If the second discrimination result is worse than the second prediction standard in predicting actual real edges as real edges and actual pseudo edges as pseudo edges, then the generative network has a good generation effect. Therefore, based on the first and second discrimination results, a target loss function can be used to update the parameters of the discriminative and generative networks. The target loss function focuses on the difference between the first discrimination result and the first predefined standard, and the difference between the second discrimination result and the second predefined standard. It should be noted that the first and second predefined standards can be the same or different.

[0078] The following section will provide a detailed explanation of how the second association graph was obtained and how the network parameters were updated during model training:

[0079] (a) Obtaining the second relationship graph:

[0080] Before perturbing the first association graph using the generative network to obtain the second association graph, the method may further include the following steps: augmenting the first association graph based on the second association prediction result to obtain an augmented graph, wherein the second association prediction result represents the edge prediction of nodes in the first association graph by the association graph neural network, and the association graph neural network is a pre-trained model obtained through machine training on the first association graph. Correspondingly, the augmented graph is perturbed using the generative network to obtain the second association graph.

[0081] First, the first association graph is augmented. Then, using the augmented graph as input, a generative network is used to obtain the second association graph. Compared to the first association graph, the augmented graph has an increased number of connected nodes, and correspondingly, the number of edges between connected nodes also increases. For example... Figure 6 As shown, a bipartite graph can be formalized as G = (U, I, E0), where U is the set of users (corresponding to multiple first sample objects), I is the set of goods (corresponding to multiple second sample objects), and E0 is the set of edges. Figure 6 The recommended edge connections correspond to the second association prediction results, and the results of increased edge connections correspond to the augmented graph.

[0082] The augmentation process for the first association graph is based on the prediction results of the second association. These second association prediction results represent the edge predictions made by the association graph neural network for nodes in the first association graph. The association graph neural network is a pre-trained model obtained through machine training on the first association graph. The association graph neural network can be a LightGCN. It's understandable that training the association graph neural network is simpler and faster than training the target graph neural network. However, to some extent, the association graph neural network performs worse than the target graph neural network in association prediction.

[0083] A graph neural network can be used to predict whether a relationship exists between any first sample object and any second sample object in a first association graph. In other words, the prediction result of the second association can indicate whether or not a relationship exists between a certain first sample object and a certain second sample object. When augmenting the first association graph, this second association prediction result, which indicates a relationship between a certain first sample object and a certain second sample object, can be used. A graph neural network can also predict the probability that a relationship exists between any first sample object and any second sample object. In other words, the prediction result of the second association can indicate the probability that a relationship exists between a certain first sample object and a certain second sample object. When augmenting the first association graph, this second association prediction result, where the probability of a relationship between a certain first sample object and a certain second sample object is greater than a probability lower bound, can be used.

[0084] Using this approach for augmentation ensures that the augmented graph, which forms the basis for perturbation, is both different from and related to the real samples (the first association graph). Based on this, the adversarial samples (the second association graph) obtained by perturbing the augmented graph are more likely to be indistinguishable from the real samples, thereby improving the association prediction performance of the trained target graph neural network.

[0085] Furthermore, the generative network can perform random pruning on the augmented graph to obtain a second association graph. The generative network can also prune the augmented graph based on a predictive pruning strategy to obtain the second association. For example... Figure 4 As shown, the generating network employs a preset pruning strategy. The step of perturbing the augmented graph using the generating network to obtain the second association graph includes:

[0086] S401: According to the preset pruning strategy, based on the product operation result of the connected nodes in the augmented graph and the attribute information of the edges between the connected nodes, determine the weight of each edge in the augmented graph, wherein the attribute information indicates whether it is a newly added edge of the augmented graph relative to the first association graph.

[0087] S402: Based on the weights and weight thresholds of each side in the augmented graph, the augmented graph is pruned to obtain the second association graph.

[0088] The preset pruning strategy here standardizes the determination of edge weights and how to prune based on these weights. The product of connected nodes can be either a Hadamard product or a dot product. For newly added edges, 1 can be used to represent the edge's attribute information. For non-new edges, 0 can be used to represent the edge's attribute information. The edge weights can be determined using an MLP for weight determination, that is, by inputting the product of connected nodes and the attribute information of the edges between those connected nodes into the MLP, the weights of the edges between the connected nodes can be quickly obtained: w u，i =MLP([z u ⊙z i ]||1 ε (e u，i )). Among them, [z u ⊙z i ] represents the result of the Hadamard product operation of connected nodes, 1 ε (e u，i ) represents the attribute information of the edge between connected nodes, u represents the first sample object in the connected nodes, i represents the second sample object in the connected nodes, and z represents the attribute information of the edge between the connected nodes. u z can represent the features of the first sample object in the connected nodes. i It can represent the features of the second sample object in the connected nodes. u It can also represent the association between the first sample object and all second sample objects in a connected node, z i It can also represent the association between the second sample object in the connected nodes and all the first sample objects.

[0089] Edge weights are used to measure the degree of correlation of the relationships represented by the edges. If the edge weight is less than the weight threshold, it means that the correlation of the relationships represented by the edges is weak, and the edges can be pruned. If the edge weight is greater than or equal to the weight threshold, it means that the correlation of the relationships represented by the edges is strong, and the edges can be retained. When determining whether to prune or retain an edge, edge weights and weight thresholds are used. The determination of edge weights takes into account both the fusion information of connected nodes (corresponding to the product operation results mentioned above) and the attribute information of the edges. The accuracy of edge weights is improved by the richer information, thereby improving the effect of pruning and obtaining adversarial examples (second correlation graph) that are "good enough to fool people".

[0090] In practical applications, when determining whether to prune or retain an edge, the Gumbel-Max technique can be used to adjust the edge weights w. u，i Processed as edge retention probability p u，i :p u，i =sigmoid(log(δ)-log(1-δ)+w u，i / τ). Where δ is a random term derived from a uniform distribution (0, 1), τ is a temperature hyperparameter, and sigmoid is a sigmoid function sigmoid(x) = 1 / 1 + e -x Correspondingly, if the probability of retaining an edge is less than the probability threshold, the edge can be pruned. If the probability of retaining an edge is greater than or equal to the probability threshold, the edge can be retained.

[0091] (ii) Parameter updates for the discriminant network and the generator network:

[0092] like Figure 5 As shown, the step of updating the parameters of the discriminator network and the generator network using a target loss function based on the first discrimination result and the second discrimination result includes:

[0093] S501: Based on the first discrimination result and the first association graph, a first loss function is constructed. The first loss function characterizes the difference between the first discrimination result and the true edge discrimination result and the false edge discrimination result in the first association graph.

[0094] S502: Based on the second discrimination result and the second association graph, construct a second loss function, which characterizes the difference between the second discrimination result and the true edge discrimination result and the false edge discrimination result in the second association graph;

[0095] S503: Construct the target loss function with the aim of maximizing the value of the first loss function and minimizing the value of the second loss function;

[0096] S504: Based on the target loss function, the parameters of the discriminant network and the parameters of the generator network are adjusted alternately.

[0097] In conjunction with the aforementioned steps S301-S305, "the target loss function focuses on the difference between the first discrimination result and the first preset standard, as well as the difference between the second discrimination result and the second preset standard."

[0098] Probability and score can be used to quantify the predicted true and false edges. In other words, if the probability and score of an edge output by the discrimination network are higher, then the edge is predicted to be a true edge to a certain extent; if the probability and score of an edge output by the discrimination network are lower, then the edge is predicted to be a false edge to a certain extent.

[0099] If the first discrimination result predicts actual real edges as real edges and actual false edges as false edges better than the first preset standard, then the discrimination network has a good discrimination effect. The first preset standard here can be applied as follows: if the probability (or score) of actual real edges is higher and the probability (or score) of actual false edges is lower in the first discrimination result, that is, the larger the difference between the probability (or score) of actual real edges and the probability (or score) of actual false edges, then the discrimination network has a good discrimination effect. Correspondingly, the larger the value of the first loss function, which characterizes the difference between the first discrimination result and the false edge discrimination result in the first association graph, the better. The first loss function can be expressed as L... BPR (f(G), y), the specific loss function can be the BPR function.

[0100] If the second discrimination result is worse than the second preset standard in predicting actual real edges as real edges and actual pseudo edges as pseudo edges, then the generation effect of the generator network is good. The second preset standard can be applied here as follows: if the probability (or score) of actual real edges is lower and the probability (or score) of actual pseudo edges is higher in the second discrimination result, that is, the smaller the difference between the probability (or score) of actual real edges and the probability (or score) of actual pseudo edges, then the generation effect of the generator network is good. Correspondingly, the smaller the value of the second loss function, which characterizes the difference between the second discrimination result and the true edge discrimination result in the second association graph, the better. The second loss function can be expressed as L... BPR (f(t(G)), y), the specific loss function can be the BPR function.

[0101] Therefore, a target loss function is constructed with the aim of maximizing the value of the first loss function and minimizing the value of the second loss function. Then, based on the target loss function, the parameters of the discriminant network and the generator network are alternately adjusted. Combining the different objectives of the discriminant and generator networks, a target loss function is constructed to alternately train the two networks. Through repeated adversarial training, the capabilities of both networks will continuously improve until a balance is reached, which is beneficial to obtaining the correlation prediction effect of the target graph neural network. Here, "alternating training" can refer to: 1. Fixing the parameters of the generator network and training the parameters of the discriminant network using the target loss function to obtain the trained discriminant network. The training objective at this time is to enable the discriminant network to distinguish between real samples and adversarial samples. 2. Fixing the parameters of the discriminant network and training the parameters of the generator network using the target loss function to obtain the trained generator network. The training objective at this time is to provide adversarial samples that are difficult for the discriminant network to distinguish. 3. Alternatingly executing the above two training steps until the training termination condition is met. The training termination condition includes, but is not limited to, the convergence of the target loss function, or the reaching of a training threshold.

[0102] Furthermore, by introducing mutual information between the first and second discrimination results: I(f(G); f(t(G))), the discriminative ability of the discriminative network and the generative ability of the generative network can be better quantified. Mutual information is a measure of the interdependence between two random variables. For discriminative networks, the greater the dependence between the first and second discrimination results, the better the discriminative network distinguishes between real samples and adversarial samples. For generative networks, the smaller the dependence between the first and second discrimination results, the better the generative network provides adversarial samples that are difficult for the discriminative network to distinguish.

[0103] Therefore, constructing the target loss function with the aim of maximizing the value of the first loss function and minimizing the value of the second loss function may include the following steps: First, constructing an estimation function representing the mutual information between the first discrimination result and the second discrimination result; then, constructing the target loss function with the aim of maximizing the value of the third loss function and minimizing the value of the fourth loss function, wherein the third loss function includes the first loss function and the estimation function: I(f(G); f(t(G)))-L BPR (f(G), y), the fourth loss function includes the second loss function and the estimation function: I(f(G); f(t(G)))+L BPR (f(t(G)), y).

[0104] The following are the specific steps for constructing an estimation function representing the mutual information between the first and second discrimination results: 1) Extract the first edge discrimination result corresponding to each node from the first discrimination result; 2) Extract the second edge discrimination result corresponding to each node from the second discrimination result; 3) For the same nodes in the first and second association graphs, determine the first correlation characterization between the first edge discrimination result and the second edge discrimination result corresponding to the same node; 4) For two dissimilar nodes in the first and second association graphs, determine the second correlation characterization between the first edge discrimination result and the second edge discrimination result corresponding to the two dissimilar nodes; 5) Obtain the estimation function based on the first correlation characterization, the second correlation characterization, and a preset processing function.

[0105] The result of the first edge judgment of a node can be regarded as the edge prediction of that node in the first association graph by the discriminant network. The result of the second edge judgment of a node can be regarded as the edge prediction of that node in the second association graph by the discriminant network. Using the discriminant network, it is possible to predict whether there is an association between any first sample object and any second sample object in the first (second) association graph. That is, the result of the first (second) edge judgment of a node can indicate whether there is an association between that node and a certain out-of-class node, or whether there is no association. The association graph neural network can also predict the probability that there is an association between any first sample object and any second sample object. That is, the result of the first (second) edge judgment of a node can indicate the probability that there is an association between that node and a certain out-of-class node.

[0106] The first and second association graphs have the same total number of nodes, so the total number of nodes can be set to B. If z i，1 z represents the result of the first edge judgment for any node in the first association graph. i，2 Let z represent the second edge discrimination result of any node in the second association graph, where i ranges from 1 to B. For the same node in the first and second association graphs, determine the first correlation degree representation between the first edge discrimination result and the second edge discrimination result corresponding to the same node, i.e., z. i，1 With z i，2 The first correlation representation between them. For two dissimilar nodes in the first and second correlation graphs, determine the second correlation representation between the first edge discrimination result and the second edge discrimination result corresponding to the two dissimilar nodes, i.e., z. i，1 With z i′，2 The second relevance representation between them. Therefore, an estimation function can be obtained based on the first relevance representations corresponding to all identical nodes, the second relevance representations corresponding to all dissimilar nodes, and a pre-defined processing function. This pre-defined processing function can utilize the idea of ​​contrastive learning to establish the relationship between these first and second relevance representations. For example, I(f(G); The sim symbol is used to calculate the relevance.

[0107] In practical applications, the generator network t can be fixed first through alternating training, and then the discriminator network f can be fixed using the following loss function: Where λ and λ reg These are hyperparameters used to control the NCE loss function and the weights of the regularization term.

[0108] As can be seen from the technical solutions provided in the embodiments of this application above, the embodiments of this application utilize a target graph neural network to obtain the first association prediction result, which improves the accuracy and efficiency of predicting the association relationship between the first target object and the second candidate object, thereby improving the accuracy of object recommendation for the first target object based on the first association prediction result. The samples used for training are association graphs. Learning the association graph is beneficial for the first association prediction result to determine whether there is an association relationship between the first target object and the second candidate object, thereby improving the object recommendation effect. At the same time, the acquisition of the target graph neural network uses adversarial training with the generator network. The generator network is responsible for providing adversarial examples. This training process is beneficial to the robustness and generalization ability of the target graph neural network model.

[0109] This application also provides a data processing apparatus, such as... Figure 7 As shown, the data processing device 70 includes:

[0110] Acquisition module 701: used to acquire a first target object and a second candidate object, wherein the second candidate object is the object to be recommended by the first target object;

[0111] Prediction module 702: used to obtain a first association prediction result by using the first target object and the second candidate object as input and a target graph neural network, wherein the first association prediction result indicates the association relationship between the first target object and the second candidate object;

[0112] Recommendation module 703: used to recommend the second candidate object to the first target object when the first association prediction result meets the preset requirements;

[0113] The target graph neural network is obtained by adversarial training of a preset graph neural network and a generator network based on a first association graph and a second association graph. The input of the preset graph neural network is the first association graph and the second association graph. The second association graph is output by the generator network based on the first association graph. The first association graph includes multiple first sample objects and multiple second sample objects as nodes and edges between connected nodes. The edges represent the association relationship between the connected nodes.

[0114] It should be noted that the apparatus and method embodiments described in the device embodiments are based on the same inventive concept.

[0115] In some embodiments, the functions or modules of the apparatus provided in this application can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0116] This application also provides a computer-readable storage medium storing at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the above-described method. The computer-readable storage medium may be a non-volatile computer-readable storage medium.

[0117] This application also provides an electronic device, which includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by the at least one processor to implement the above method.

[0118] Electronic devices can be provided as terminals, servers, or other forms of devices.

[0119] Figure 8 A block diagram of an electronic device according to an embodiment of this application is shown. For example, electronic device 1900 may be provided as a server. (Refer to...) Figure 8 The electronic device 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by memory 1932 for storing instructions, such as application programs, that can be executed by the processing component 1922. The application programs stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 1922 is configured to execute instructions to perform the methods described above.

[0120] Electronic device 1900 may also include a power supply component 1926 configured to perform power management of electronic device 1900, a wired or wireless network interface 1950 configured to connect electronic device 1900 to a network, and an input / output (I / O) interface 1958. Electronic device 1900 can operate on an operating system stored in memory 1932, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, or similar.

[0121] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of an electronic device 1900 to perform the above-described method.

[0122] This application may be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this application.

[0123] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0124] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0125] The computer program instructions used to perform the operations of this application may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C+, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuits, such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), are personalized by utilizing the status information of the computer-readable program instructions. These electronic circuits can execute the computer-readable program instructions to implement various aspects of this application.

[0126] Various aspects of this application are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0127] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0128] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0129] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction, which includes one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions specified in the blocks may occur in a different order than those specified in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0130] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technological improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A data processing method, characterized in that, The method includes: Obtain a first target object and a second candidate object. The second candidate object is the object to be recommended by the first target object. The first target object is an account object, and the second candidate object is a multimedia object or a product object. Using the first target object and the second candidate object as input, a first association prediction result is obtained using a target graph neural network. The first association prediction result indicates the association between the first target object and the second candidate object. The association represents whether the first target object performs at least one of the following interactive behaviors on the second candidate object: clicking, browsing, listening, watching, collecting, following, downloading, forwarding, and sharing. If the first correlation prediction result meets the preset requirements, the second candidate object is recommended to the first target object; The target graph neural network is obtained through the following training method: A first association graph is obtained, comprising multiple first sample objects and multiple second sample objects as nodes, and edges between connected nodes, wherein the edges represent the association relationship between the connected nodes; a generator network is used to perturb the first association graph to obtain a second association graph; a discriminator network is used to perform edge discrimination on the first and second association graphs respectively, obtaining a first discrimination result corresponding to the first association graph and a second discrimination result corresponding to the second association graph; based on the first and second discrimination results, a target loss function is used to update the parameters of the discriminator network and the generator network; the steps of obtaining the second association graph and updating the parameters of the discriminator network and the generator network are repeated until the value of the target loss function meets the convergence condition, and the discriminator network corresponding to the current parameters is determined to be the target graph neural network.

2. The method according to claim 1, characterized in that, Before perturbing the first association graph using the generator network to obtain the second association graph, the method further includes: The first association graph is augmented based on the second association prediction result to obtain an augmented graph. The second association prediction result represents the association graph neural network's prediction of the edges connecting nodes in the first association graph. The association graph neural network is a pre-trained model obtained by machine training on the first association graph. The step of perturbing the first association graph using the generating network to obtain the second association graph includes: perturbing the augmented graph using the generating network to obtain the second association graph.

3. The method according to claim 2, characterized in that, The generating network employs a preset pruning strategy. The step of perturbing the augmented graph using the generating network to obtain the second association graph includes: According to the preset pruning strategy, based on the product operation result of the connected nodes in the augmented graph and the attribute information of the edges between the connected nodes, the weight of each edge in the augmented graph is determined, and the attribute information indicates whether it is a new edge of the augmented graph relative to the first association graph. Based on the weights and weight thresholds of each side in the augmented graph, the augmented graph is pruned to obtain the second association graph.

4. The method according to claim 1, characterized in that, The step of updating the parameters of the discriminator network and the generator network based on the first discrimination result and the second discrimination result using a target loss function includes: Based on the first discrimination result and the first association graph, a first loss function is constructed. The first loss function characterizes the difference between the first discrimination result and the true edge discrimination result and the false edge discrimination result in the first association graph. Based on the second discrimination result and the second association graph, a second loss function is constructed. The second loss function characterizes the difference between the second discrimination result and the true edge discrimination result and the false edge discrimination result in the second association graph. The target loss function is constructed with the aim of maximizing the value of the first loss function and minimizing the value of the second loss function. Based on the target loss function, the parameters of the discriminator network and the generator network are adjusted alternately.

5. The method according to claim 4, characterized in that, The construction of the target loss function, aimed at maximizing the value of the first loss function and minimizing the value of the second loss function, includes: Construct an estimation function that represents the mutual information between the first discrimination result and the second discrimination result; The target loss function is constructed with the aim of maximizing the value of the third loss function and minimizing the value of the fourth loss function. The third loss function includes the first loss function and the estimation function, and the fourth loss function includes the second loss function and the estimation function.

6. The method according to claim 5, characterized in that, The construction of the estimation function representing the mutual information between the first discrimination result and the second discrimination result includes: Extract the first edge discrimination result corresponding to each node from the first discrimination result; extract the second edge discrimination result corresponding to each node from the second discrimination result; For the same nodes in the first association graph and the second association graph, a first correlation characterization is determined between the first edge discrimination result and the second edge discrimination result corresponding to the same node; For two dissimilar nodes in the first and second association graphs, a second correlation characterization is determined between the first edge discrimination result and the second edge discrimination result corresponding to the two dissimilar nodes; The estimation function is obtained based on the first correlation characterization, the second correlation characterization, and the preset processing function.

7. A data processing apparatus, characterized in that, The device includes: Acquisition module: used to acquire a first target object and a second candidate object, wherein the second candidate object is the object to be recommended by the first target object, the first target object is an account object, and the second candidate object is a multimedia object or a product object; Prediction module: used to obtain a first association prediction result by using the first target object and the second candidate object as input and a target graph neural network. The first association prediction result indicates the association between the first target object and the second candidate object. The association characterizes whether the first target object performs at least one of the following interactive behaviors on the second candidate object: click, browse, listen, watch, favorite, follow, download, forward, share. Recommendation module: used to recommend the second candidate object to the first target object when the prediction result of the first association relationship meets the preset requirements; The target graph neural network is obtained through the following training method: A first association graph is obtained, comprising multiple first sample objects and multiple second sample objects as nodes, and edges between connected nodes, wherein the edges represent the association relationship between the connected nodes; a generator network is used to perturb the first association graph to obtain a second association graph; a discriminator network is used to perform edge discrimination on the first and second association graphs respectively, obtaining a first discrimination result corresponding to the first association graph and a second discrimination result corresponding to the second association graph; based on the first and second discrimination results, a target loss function is used to update the parameters of the discriminator network and the generator network; the steps of obtaining the second association graph and updating the parameters of the discriminator network and the generator network are repeated until the value of the target loss function meets the convergence condition, and the discriminator network corresponding to the current parameters is determined to be the target graph neural network.

8. The apparatus according to claim 7, characterized in that, Before perturbating the first association graph using the generative network to obtain the second association graph, the first association graph is augmented based on the second association prediction result to obtain an augmented graph. The second association prediction result represents the association graph neural network's prediction of the edges connecting nodes in the first association graph. The association graph neural network is a pre-trained model obtained through machine training on the first association graph. The step of perturbing the first association graph using the generating network to obtain the second association graph includes: perturbing the augmented graph using the generating network to obtain the second association graph.

9. The apparatus according to claim 8, characterized in that, The generating network employs a preset pruning strategy. The step of perturbing the augmented graph using the generating network to obtain the second association graph includes: According to the preset pruning strategy, based on the product operation result of the connected nodes in the augmented graph and the attribute information of the edges between the connected nodes, the weight of each edge in the augmented graph is determined, and the attribute information indicates whether it is a new edge of the augmented graph relative to the first association graph. Based on the weights and weight thresholds of each side in the augmented graph, the augmented graph is pruned to obtain the second association graph.

10. The apparatus according to claim 7, characterized in that, The step of updating the parameters of the discriminator network and the generator network based on the first discrimination result and the second discrimination result using a target loss function includes: Based on the first discrimination result and the first association graph, a first loss function is constructed. The first loss function characterizes the difference between the first discrimination result and the true edge discrimination result and the false edge discrimination result in the first association graph. Based on the second discrimination result and the second association graph, a second loss function is constructed. The second loss function characterizes the difference between the second discrimination result and the true edge discrimination result and the false edge discrimination result in the second association graph. The target loss function is constructed with the aim of maximizing the value of the first loss function and minimizing the value of the second loss function. Based on the target loss function, the parameters of the discriminator network and the generator network are adjusted alternately.

11. The apparatus according to claim 10, characterized in that, The construction of the target loss function, aimed at maximizing the value of the first loss function and minimizing the value of the second loss function, includes: Construct an estimation function that represents the mutual information between the first discrimination result and the second discrimination result; The target loss function is constructed with the aim of maximizing the value of the third loss function and minimizing the value of the fourth loss function. The third loss function includes the first loss function and the estimation function, and the fourth loss function includes the second loss function and the estimation function.

12. The apparatus according to claim 11, characterized in that, The construction of the estimation function representing the mutual information between the first discrimination result and the second discrimination result includes: Extract the first edge discrimination result corresponding to each node from the first discrimination result; extract the second edge discrimination result corresponding to each node from the second discrimination result; For the same nodes in the first association graph and the second association graph, a first correlation characterization is determined between the first edge discrimination result and the second edge discrimination result corresponding to the same node; For two dissimilar nodes in the first and second association graphs, a second correlation characterization is determined between the first edge discrimination result and the second edge discrimination result corresponding to the two dissimilar nodes; The estimation function is obtained based on the first correlation characterization, the second correlation characterization, and the preset processing function.

13. An electronic device, characterized in that, The electronic device includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores at least one instruction or at least one program, the at least one instruction or at least one program being loaded and executed by the at least one processor to implement the data processing method as described in any one of claims 1-6.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction or at least one program, which is loaded and executed by a processor to implement the data processing method as described in any one of claims 1-6.

15. A computer program product, characterized in that, The computer program product includes at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the data processing method as described in any one of claims 1-6.

Images